Re: [PATCH 1/2] vfs: Do not allow mnt_longterm to go negative

[Al Viro <viro@ZenIV.linux.org.uk>]

On Tue, May 29, 2012 at 12:56:44PM +0200, Luk?? Czerner wrote:
> 
> Anyone can take a look at this ? It's pretty pressing I would say
> since this bug will not only disallow to unattach the loop devices
> but also make it unable to actually unmount the file system!

I agree that it's a bug, but I don't believe that this is a proper
fix.  We can do it much simpler.  Look: making these guys longterm
happens exactly at the same points where we set their ->mnt_ns non-NULL.
Remember that "longterm" == "internal kernel mount, mount in some
namespace, someone's root or pwd".  So the problem is that umount_tree()
is that umount_tree() blindly assumes that it's killing a longterm
ref, even though it may be called when dissolving a tree that never
made it into anyone's namespace.  Fine, so let's just have that
sucker do
                if (p->mnt_ns)
                        __mnt_make_shortterm(p);
                p->mnt_ns = NULL;
instead and that's all it takes.  Alternatively, we can just turn the
code in mntput_no_expire() into
	if (likely(mnt->mnt_ns || atomic_read(...))
and get rid of longterm/shortterm logics around ->mnt_ns completely.
Short-term, pardon the bad pun, I'm going with the first variant; it's
less intrusive.  Longer term I think we want to get rid of playing with
mnt_longterm in there and just check ->mnt_ns as well when deciding to
go for the fast path in mntput_no_expire().

Fixed in my tree, will push in a couple of hours.  And that'll need to
go into -stable as well.  Thanks for catching that sucker...